The gap it closes
Most teams can now generate an AI application faster than their organisation can govern, cost-control or deliver it. The usual answer is five products bolted together: a model gateway, a cost dashboard, an app builder, an access-control layer and a deployment story.
DBS Invenio is one system, so the application is metered, permissioned, audited and deliverable the moment it exists. It runs on your own infrastructure, on-premises and air-gapped if that is what your site requires.
A harness, not a wrapper around a model
Calling a model is the easy part. The harness is everything wrapped around that call, and it is the part we built first:
- A loop that knows when to stop. Tools are called in a bounded loop with an iteration ceiling and a per-connector circuit breaker. A confused agent stops and says so instead of spending your budget.
- Memory that carries forward. Conversation memory is extracted into durable fragments, organised into topics and injected into later runs, so the next task does not start from nothing.
- It can stop and ask you. An agent can pause mid-task, put a real question to a person with options to choose from, and carry on from where it paused when the answer comes back. No restart, no lost context.
- Its output is checked before it counts. Structured work — a workflow, a function, an application — is validated first, and the validator's own errors are handed back to the agent to fix. Invalid work never reaches a saved version.
- Retry that knows when not to. Rate limits and overloads back off with jitter and honour the provider's
Retry-After. Mid-stream failures are never retried, because a retry there would duplicate rather than recover.
Build it by describing it
An agent builds a complete, working web application from a plain-language description, and emits it as portable HTML and JavaScript you can read, diff and host anywhere — not a proprietary metamodel you can never leave.
From there it goes where the people are: published to the web, dropped into an existing application as a widget, wrapped as a desktop client, or synced to a phone that then runs it with no network at all, with camera, GPS, NFC and biometrics available to it as governed tools.
Nothing an agent writes goes live on its own
AI-generated software gets the controls real software gets: immutable versioned snapshots, an explicit human promote that also freezes the application's runtime configuration, capability keys enforced in the resolver rather than only in the interface, and a sandboxed JavaScript runtime with no host type access.
One design decision is worth stating plainly. No tool surface anywhere in the platform exposes a promote. An agent that has been prompt-injected can edit a development version and still cannot publish it.
Every token attributed, every budget failing closed
One gateway covers Claude, OpenAI, Azure and any OpenAI-compatible endpoint, including Ollama and LM Studio at zero marginal cost, reconfigured from the database without a restart. Every round-trip is attributed to a user, a project and an application, and a published application can carry a daily token budget and a rate limit that refuse the call rather than warn you afterwards.
What it deliberately does not do
It does not score the quality of model output: there are no golden sets and no regression testing on answers. It does not inspect prompts or completions for content either. The answer to that second one is architectural rather than a filter — what an agent writes lands on a development version, and a person decides whether it goes live.
Where it fits with the rest of what we do
Invenio is the layer above your systems of record, not a replacement for them. Your typed data model stays in the system that already defines and enforces it, and Invenio connects to it. The AI assistants in DynamicPlantMesh run on Invenio, which is the same pattern applied to plant data: the historian stays the system of record, and the harness is what answers questions about it.
Find out more
The full product site — platform detail, demos, comparisons, the security and trust pages and pricing — is at dbsinvenio.com. If you would rather talk it through, tell us what you are trying to build and we will reply within one working day.